Five AI video mistakes I made so you don't have to
Wobbly hands, melting faces, and too many ideas in one shot.
My first AI video was supposed to be a ten second clip of a man pouring tea. What I got: a man with seven fingers pouring tea upward into the ceiling. Fun, but not usable.
Here's what I learned since.
1. One action per clip. Pouring tea. Not pouring tea while walking and talking and turning around.
2. Describe the camera. "Slow push-in, eye level, 35mm" gives the model something to hold on to. Without it, the camera drifts around like it's drunk.
3. Keep hands out of frame if you can. Models are better at hands now, but they still slip.
4. Lighting matters more than detail. "Soft morning window light" fixes more than "ultra detailed 8K".
5. Shorter is better. Four seconds that look real beat twelve seconds that fall apart at second nine.
Then stitch the good clips together in an editor. That's how most of the nice AI videos you see online are made, really. Lots of short clips, lots of rejects.